Dynamic Jump Intensities and Risk Premiums in Crude Oil Futures and Options Markets

Transcription

1 Dynamic Jump Intensities and Risk Premiums in Crude Oil Futures and Options Markets Peter Christoffersen University of Toronto, CBS, and CREATES Kris Jacobs Bingxin Li University of Houston and Tilburg University University of Houston January 15, 213 Work in Progress Please do not cite Abstract Options on crude oil futures are the most actively traded commodity derivatives. Existing pricing models for crude oil derivatives are computationally intensive due to the presence of latent state variables. We adopt a class of computationally efficient discrete-time jump models that allow for closed-form option valuation, and investigate the economic importance of jumps and dynamic jump intensities in the market for crude oil futures and futures options. Including jumps is crucial for modeling crude oil futures and futures options, and we find very strong evidence in favor of time-varying jump intensities. The main role of jumps and jump risk in the crude oil futures and options markets is to capture excess kurtosis in the data. We find that jumps account for a large part of the variation in crude oil futures and options prices, and a substantial part of the risk premium is due to jumps. Futures data indicate the presence of many small jumps, while option data point towards large infrequent jumps. JEL Classification: G12; G13. Keywords: Crude oil; Futures; Options; Discrete-time models; Jump intensities; Risk premiums. 1

2 1 Introduction Crude oil is the single largest commodity in international trade, and the crude oil derivatives market constitutes the most liquid commodity derivatives market. In December 211, WTI and Brent crude oil futures accounted for 51.4% of dollar weight in the S&P GSCI commodity index. Nearly 9, futures and options contracts trade on WTI every day. In order to price and hedge this increasingly important commodity, it is crucial to model crude oil futures and options and better understand their dynamics. Surprisingly though, there are relatively few studies on pricing crude oil derivatives, especially when compared with the existing literature on equity derivatives. Trolle and Schwartz (29) estimate a continuous-time stochastic volatility model using NYMEX crude oil futures and options and find evidence for two, predominantly unspanned, volatility factors. They do not consider jump processes, which have been used in other security markets to model large movements. Larsson and Nossman (21) examine the performance of affine jump diffusion models with stochastic volatility for modeling the time series of crude oil spot prices. Their results show that stochastic volatility alone is not sufficient and jumps are an essential factor to correctly capture the time series properties of oil prices. However, they do not use panel data on futures contracts nor option prices. Hamilton and Wu (211) model crude oil futures with an affine term structure model and document significant changes in oil futures risk premia since 25. Pan (211) studies the relationship of state price densities and investor beliefs. Chiarella, Kang, Sklibosios, and To (212) document a hump-shaped volatility structure in the commodity derivatives market. To the best of our knowledge, no existing studies have implemented jump models using extensive cross-sections of crude oil derivatives. There is an extensive literature on the modeling of equity derivatives, which mainly focuses on index returns and options. This literature includes models with both stochastic volatility and jumps (e.g., see Bakshi, Cao, and Chen, 1997; Bates, 1996, 2, 26; Pan 22; Eraker, 24; Carr and Wu, 24; Santa-Clara and Yan 21,), as well as GARCH-style jump models (Maheu and McCurdy, 24; Duan, Ritchken and Sun, 26, Christoffersen, Jacobs, and Ornthanalai, 212; and Ornthanalai, 212). The implementation of continuous-time stochastic volatility models with Poisson jumps is complex, because the likelihood function is typically not available in closed form, and therefore option pricing in the presence of jumps typically relies on complex econometric methods to filter the unobserved state variables. For example, Chernov, Gallant, Ghysels, and Tauchen (23) use an Efficient Method of 2

3 Moments (EMM) based method, Pan (22) uses the implied-state Generalized Method of Moments (GMM) technique, Eraker, Johannes, and Polson (23), Eraker (24), and Li, Wells, and Yu (27) employ Markov Chain Monte Carlo (MCMC) techniques, and Trolle and Schwartz (29) use the Extended Kalman Filter (EKF). This type of estimation is computationally intensive, especially when dealing with large data sets. In this paper, we study discrete-time models in which the conditional variance of the normal innovation and the conditional jump intensity of a compound Poisson process are governed by GARCHtype dynamics. We estimate four nested jumps models to investigate the importance of time-varying conditional variance and time-varying jump intensity using an extensive panel data set of crude oil futures and option prices. For the discrete-time models in this paper, both the conditional jump intensity and the conditional variance can be directly computed from the observed shocks using an analytical filter. With the analytical filter, filtering the normal component and the jump component is relatively simple and extremely fast, even when the jump intensity is time-varying. It takes less than a second to filter 38,24 futures contracts using Matlab on a standard PC. Because the variance and the jump intensity dynamic can be updated analytically, we can conveniently estimate the model using MLE or NLS estimation. Calculating the Implied Volatility Root Mean Squared Error (IVRMSE) of 283,653 option contracts takes about 7 seconds. We investigate the economic importance of jumps and dynamic jump intensity in the crude oil market and compare the fit of the jump models with that of a benchmark GARCH model without jumps. We study four jump models with different specifications of jump intensity and conditional variance. The DVCJ (dynamic volatility with constant jumps) model is assumed to have constant jump intensities, which is consistent with most of the existing continuous-time SVJ literature. The CVDJ (constant volatility and dynamic jumps) model has time-varying jump intensity, but the normal innovation to the return process is assumed to be homoskedastic. It is related to the constant volatility jump-diffusion models of Bates (1991) and Merton (1976). The DVDJ (dynamic volatility and dynamic jumps) model contains time-varying jump intensities and a time-varying conditional variance but the dynamics are driven by the same dynamic. The DVSDJ (dynamic volatility with separate dynamic jumps) model is the most general model in which both the jump intensity and the conditional variance are time-varying and the dynamics are modeled separately. The latter two models are related to the most complex SVJ dynamics of Eraker (24) and Santa-Clara and Yan (21). 3

4 We find strong evidence of the presence of jumps and dynamic jump intensities in the crude oil market. During crisis periods, when market risk is high, jumps occur more frequently. Jump models with time-varying jump intensities outperform the benchmark model significantly, and in models with timevarying intensities jumps explain a large part of the variance of the underlying futures data, regardless of the data used in estimation. Estimates based on futures data indicate the presence of many small jumps, whereas estimates based on option data point towards infrequent large jumps. The implied conditional kurtosis from option-based estimates is much larger than that implied by futures data. Contrary to equity index markets, the main role of jumps and jump risk in crude oil futures markets is to capture the excess kurtosis in the data, while skewness is of second-order importance. The rest of the paper proceeds as follows. Section 2 develops discrete-time pricing models for commodity futures. We propose four nested jump models and a benchmark GARCH model. Section 3 discusses the crude oil futures and options data used in this paper. In Section 4 we explain MLE estimation on futures contracts and report the estimation results. Section 5 derives the risk-neutral dynamics and the closed form option valuation formula. Section 6 presents estimation results using options data, and Section 7 presents results using options and futures jointly in estimation. Section 8 concludes. 2 Models for Commodity Futures Markets In commodity futures markets, we observe futures prices for different maturities, and spot prices are considered unobservable. We develop a general class of discrete-time models by first specifying models for the unobservable spot price, and then imposing the relationship between the spot price and futures prices for different maturities. The relationship between the futures price and the spot price of a commodity is given by F(t, T) = S(t)exp [c(t, T) (T t)] (2.1) where S(t) denotes the spot price at time t, F(t, T) is the time t price of the futures contract maturing at time T, and c(t, T) denotes the time t forward cost of carry for the futures contract maturing at T. Combining this with the expression for the time t+1 futures price maturing at time T, the log futures return is given by 4

5 log F t+1,t = log S t+1 + c(t, t + 1, T) (2.2) F t,t S t where c(t, t + 1, T) captures the cumulative difference of the forward cost of carry between time t and t+1 for futures contracts maturing at time T. The forward cost of carry c(t, T) is unobservable. We assume it is deterministic, and extract it from the relationship between futures and spots prices following the approach of Gibson and Schwartz (199). We provide details on this method in Appendix A. Several studies have estimated more general models with stochastic cost of carry specifications. We implemented stochastic cost of carry models, but for the models under study here the cost of carry model does not seem to substantially affect the results. 2.1 The Benchmark Model We formulate a new class of jump models for commodities markets. To provide a benchmark for these models that can capture several important stylized facts of commodity markets, we first consider a standard GARCH model. We specify the spot return process log S t+1 S t = r t+1 + λ z 1 2 h z,t+1 + z t+1 (2.3) where r t+1 is the instantaneous risk free rate, z t+1 is an innovation which is distributed N, h z,t+1, h z,t+1 is the conditional variance known at time t, and λ z is the market price of risk associated with the normal innovation. Substituting (2.3) into (2.2), we get the benchmark GARCH model for futures returns log F t+1,t F t,t = r t+1 + λ z 1 2 h z,t+1 + h z,t+1 ε z,t+1 + c(t, t + 1, T) (2.4) where ε z,t+1 is distributed N(,1). The conditional variance of the normal innovation h z,t+1 is governed by a GARCH (1, 1) process, which is specified according to Heston and Nandi (2). h z,t+1 = ω z + b z h z,t + a z (ε z,t c z h t ) 2 (2.5) GARCH models provide a convenient framework to capture stylized facts in financial markets such as conditional heteroskedasticity, volatility clustering, and mean reversion in volatility. These stylized facts are also very prominent in commodity futures markets. The GARCH dynamic in (2.5) is different from the more conventional GARCH specifications of Engle (1982) and Bollerslev (1986), and is explicitly 5

6 designed to facilitate option valuation. We discuss the benefits of the specification in (2.5) in more detail below. Consistent with other GARCH specifications, the conditional variance h z,t+1 in (2.5) is predictable conditional on information available at time t. The unconditional variance is given by E h z,t+1 = (ω z + a z )/(1 b z a z c 2 2 z ), where b z + a z c z is the variance persistence. Further, given a positive estimate for a z, the sign of c z determines the correlation between the futures returns and the conditional variance. Equivalently, c z can be thought of as controlling the skewness or asymmetry of the distribution of log returns, with a positive c z resulting in a negatively skewed multi-day distribution. 2.2 Filtering the Conditional Variance Using Futures with Different Maturities Heston and Nandi (2) estimate their model using S&P 5 index returns and options. For that application, it is straightforward to directly filter the conditional variance h z,t+1 from the return innovations with the GARCH model. Our application is more complex, because we have a cross-section of crude oil futures prices with different maturities (eight time series of futures contracts in our case). Consider the following futures return series log F t+1,t i F t,ti = r t+1 + λ i,z 1 2 h i,z,t+1 + h i,z,t+1 ε i,z,t+1 + c(t, t + 1, T i ) (2.6) where i = 1, 2,..,8 and T i represents maturity date of the futures contracts maturing in i month. The number of parameters in this model is very large if the parameters for the return and variance dynamics are maturity-specific, and the number of parameters increases with the number of futures contracts. Instead, impose the restriction that these parameters are the same for all eight maturities. Furthermore,instead of separate dynamics for each maturity, consider filtering the volatility using h z,t+1 = ω z + b z h z,t + a z ε 1,z,t c z h z,t 2 + a z ε 2,z,t c z h z,t a z ε 8,z,t c z h z,t 2 (2.7) Finally, we have to specify the structure of the covariance matrix of the vector of residuals. To keep the model as parsimonious as possible and further reduce the number of parameters, we assume that covariance matrix of the return innovations is a diagonal matrix, ε z,t+1 = ε 1,z,t+1, ε 2,z,t+1,, 6

7 ε 8,z,t+1 ~ i. i. d. N(, I 8 ), where I 8 is the identity matrix. This approach is consistent with the assumptions made by Trolle and Schwartz (29), who use the Kalman filter for estimation. 2.3 Commodity Futures Returns with Dynamic Jump Intensities The futures return process in (2.4)-(2.5) provides a benchmark model that can capture several important stylized facts using a simple setup with a single normal innovation. We now provide a much richer class of models with jumps in returns and volatilities, and with potentially time-varying jump intensities. Futures returns are given by log F t+1,t F t,t = r t+1 + λ z 1 2 h z,t+1 + λ y ξ h y,t+1 + z t+1 + y t+1 + c(t, t + 1, T). (2.8) where the notation for r t+1, λ z, z t+1 and h z,t+1 is the same as in Section 2.1. The jump component y t+1 is specified as a Compound Poisson process denoted as J h y,t+1, θ, δδ 2. The Compound Poisson structure assumes that the jump size is independently drawn from a normal distribution with mean θ and variance δδ 2. The number of jumps n t+1 arriving between times t and t+1 is a Poisson counting process with intensity h y,t+1. The jump component in period t+1 is therefore given by n t+1 j y t+1 = x t+1 j= (2.9) j where x t+1, j =,1, 2, is an i.i.d. sequence of normally distributed random variables, x t+1 ~N(θ, δδ 2 ). The conditional expectation of the number of jumps arriving over time interval (t, t+1) equals the jump intensity, E t [n t+1 ] = h y,t+1. The conditional mean and variance of the jump component y t+1 are given by θh y,t+1 and (θ 2 + δδ 2 )h y,t+1 respectively. j The convexity adjustment terms 1 2 h z,t+1 and ξh y,t+1 (e θ+δ2 2 1)h y,t+1 in (2.6) act as compensators to the normal and jump component respectively. The conditional risk premium is given by γ t+1 λ z h z,t+1 + λ y h y,t+1, with λ z and λ y denoting the market prices of risks for the normal and jump components. 2.4 Four Jump Models 7

8 We formulate four nested models with jump dynamics. In the most general model, we assume that both the conditional variance of the normal component and the jump intensity are governed by the following extended GARCH (1, 1) dynamics. h z,t+1 = ω z + b z h z,t + a z h z,t (z t c z h z,t ) 2 + d z y t (2.1) h y,t+1 = ω y + b y h y,t + a y h z,t (z t c y h z,t ) 2 + d y y t (2.11) where ω z, a z, b z, c z, d z, ω y, a y, b y, c y, d y are parameters to be estimated. We refer to this specification as the Dynamic Volatility and Separate Dynamic Jumps (DVSDJ) model. It is a very rich and flexible model. First, it allows for jumps in volatility as well as jumps in returns. It has been shown in the index option literature that jumps in volatility are useful to explain option volatility smiles and smirks (see for example Eraker, Johannes and Polson, 23; and Eraker, 24). Moreover, in (2.1)-(2.11), the normal and jump innovations, z t and y t, enter separately into the GARCH updating dynamics. The model therefore allows each type of innovation to impact the variance and jump intensity separately. Second, the model is designed to yield closed-form expressions for securities prices. Critically, we are able to derive analytical results for option valuation. The specification of h z,t+1 and h y,t+1 in (2.1)-(2.11) therefore has substantial advantages. One potential problem is that the model is richly parameterized. This presumably will not create problems for in-sample option valuation, but it may cause the model s out-of-sample performance to deteriorate. We therefore also investigate three nested specifications, which impose restrictions on h z,t+1 and h y,t+1, and greatly reduce the dimension of the parameter space. The first nested model is the Dynamic Volatility and Constant Jumps (DVCJ) model, which imposes the following restrictions b y =, a y =, c y =, d y = (2.12) This model maintains the normal component s GARCH dynamic, but jumps arrive at a constant rate ω y, regardless of the level of volatility in the market. The second nested model is the Constant Volatility and Dynamic Jumps (CVDJ) model, which imposes the restrictions b z =, a z =, c z =, d z = (2.13) 8

9 In this model, time-variation in the return distribution is driven by the jump component only. The normal component of returns is homoskedastic, with a constant variance equal to ω z. In the third nested model, the Dynamic Volatility and Dynamic Jumps (DVDJ) model, h z,t+1 and h y,t+1 are both time-varying but driven by the same dynamic. The conditional jump intensity is affine in the conditional variance of the normal component h y,t+1 = kh z,t+1 (2.14) where k is a parameter to be estimated. The DVDJ specification can be written as a special case of the DVSDJ model, subject to the following restrictions on the parameters of h y,t+1 ω y = ω z k, b y = b z, a y = a z k, c y = c z, d y = d z k (2.15) The specification of these four models is based on the index returns models in Ornthanalai (212) and Christoffersen, Jacobs, and Ornthanalai (CJO, 212). We follow the classification of jump models proposed in CJO (212), but the model dynamics in (2.1)-(2.11) are different in a very important way. They are designed to yield a closed-form solution for option prices, and in order to do so we have adopted a rather simple specification for jumps in volatility. The drawback of this approach is that we cannot ensure positive volatility. 3. Crude Oil Futures and Options Data We now discuss the crude oil futures and options data used in the empirical analysis, and present summary statistics. We use a data set of Chicago Mercantile Exchange (CME group, formerly NYMEX) crude oil futures and options data. We use a sample of daily data from January 2 nd, 199 to December 3 rd, 28. The CME crude oil derivatives market is the world s largest and most liquid commodity derivatives market. The range of maturities covered by futures and options and the range of option strike prices are also greater than for other commodities (for a discussion see Trolle and Schwartz, 29, henceforth TS). We screen futures contracts based on patterns in trading activity. Open interest for futures contract tends to peak approximately two weeks before expiration. Among futures and options with more than two weeks to expiration, the first six monthly contracts tend to be very liquid. For contracts with maturities over six months, trading activity is concentrated in the contracts expiring in March, June, September, and December. 9

10 Following TS (29), we therefore screen the available futures and options data according to the following procedure: discard all futures contracts with 14 or less days to expiration. Among the remaining, retain the first six monthly contracts. Furthermore, choose the first two contracts with expiration either in March, June, September or December. This procedure leaves us with eight futures contract series which we label M1, M2, M3, M4, M5, M6, Q1, and Q2. We include the following options on these eight futures contracts. For each option maturity, we consider eleven moneyness intervals: , ,.86-.9,.9-.94, , , , , , , and Moneyness is defined as option strike divided by the price of the underlying futures contract. Among the options within a given moneyness interval, we select the one that is closest to the mean of the interval. Our data consist of American options on crude oil futures contracts. 1 CME has also introduced European-style crude oil options, which are easier to analyze. However, the trading history is much shorter and liquidity is much lower than for the American options. Since the pricing formulae are designed for European options, we have to convert the American option prices to European option prices. We assume that the price of the underlying futures contract follows a geometric Brownian motion. This specification allows for very accurate pricing of the American option using the Barone-Adesi and Whaley (1987) formula. Inverting this formula yields a log-normal implied volatility, from which we can subsequently obtain the European option price using the Black (1976) formula. To minimize the effect of errors in the early exercise approximation, we use only OTM and ATM options, i.e., puts with moneyness less than one and calls with moneyness greater than one. In addition, we only consider options that have open interest in excess of 1 contracts and options with prices larger than.1 dollars. This data filtering procedure yields 38,24 futures contracts and 283,653 option contracts observed over 4,753 business days. The number of futures contracts is 8 on every day of the sample, while the number of option contracts is between 23 and 87. Figure 1 displays the prices of the futures contracts. All prices in this paper are settlement prices. 2 To avoid cluttering the figure, we only display the futures term structure on Wednesdays. Futures prices increase dramatically between 23 and 27, and subsequently decline. Consistent with existing studies (Trolle and Schwartz (29), Litzenberger and Rabinowitz (1995)), we find that the crude oil market is on 1 Futures contracts expire on the third business day prior to the 25 th calendar day (or the business day right before it if the 25 th is not a business day) of the month that precedes the delivery month. Options written on futures expire three business days prior to the expiration date of the futures. 2 The CME light, sweet crude oil futures contract trades in units of 1 barrels. Prices are quoted in US dollars per barrel. 1

11 average in backwardation. Also, the price of long maturity futures contracts such as Q2 is lower than that of short maturity futures contracts. Figure 2 plots the daily returns, log F t+1,t F t,t, for the eight (M1, M2, M3, M4, M5, M6, Q1, Q2) futures contracts and Panel A of Table 1 provides summary statistics. Table 1 indicates that futures returns on longer maturities futures contracts, e.g., Q2 futures contracts, are less volatile than futures returns for shorter maturity contracts. However, Figure 2 indicates that returns of futures contracts with long maturity seem to have more spikes, which can be interpreted as jumps. Table 1 also reports summary statistics for higher moments of the daily futures returns. On average across maturities, skewness is -.91 and kurtosis is The daily crude oil futures return series is thus skewed towards the left, indicating that there are more negative than positive outlying returns in the crude oil market. Moreover, the return series is characterized by a distribution with tails that are significantly thicker than a normal distribution. Panel B of Table 1 lists the average number of option contracts across maturity and moneyness. The number of option contracts decreases with maturity. Among the 11 moneyness intervals, the number of option contracts is highest in the ATM interval. Panel C reports the average option prices. As expected, the average price of the option contracts increases as the maturity of underlying futures contracts increases. Figure 3 displays the implied ATM log-normal volatilities, and Panel D of Table 1 reports their averages by moneyness and maturity. Options with short maturities have higher implied volatilities than options with long maturities. Large spikes in the option implied volatilities appear around the end of 199 and beginning of 1991, at the time of the first Gulf War, around the September 21 terrorist attack, the second Gulf War in March 23, and during the financial crisis in 28. Among the eleven moneyness intervals, the average implied volatilities are lowest for ATM options. The data exhibit a smirk for several maturities, but for other maturities it is not clear if the smirk pattern is economically significant, or if the data are instead characterized by a smile. These patterns are important with respect to the relative role of return skewness and kurtosis for characterizing the option data. Panel D of Table 2 suggests that modeling kurtosis may be more critical than capturing skewness for crude oil data. This is consistent with the descriptive statistics for the futures data in Table Evidence from Futures Prices 11

12 We first discuss how to use maximum likelihood to estimate the models using futures returns. We then present parameter estimates for the four proposed models as well as the benchmark GARCH model. Subsequently we use the parameter estimates to investigate the models most important implications for option valuation. 4.1 Maximum Likelihood Estimation using Futures Data We estimate the model parameters using Maximum Likelihood (MLE). The likelihood function for returns depends on the normal and Compound Poisson distributions. The conditional density of the ith returns process in equation (2.8), given that there are n t+1 = j jumps occurring between period t and t+1, is given by 1 f t R i,t+1 n t+1 = j = exp R i,t+1 μ i,t+1 jθ 2. 2π h z,t+1 + jδδ 2 2 h z,t+1 + jδδ 2 (4.1) where R i,t+1 log F t+1,t i F t,ti, and μ i,t+1 = r t+1 + λ z 1 2 h z,t+1 + λ y ξ h y,t+1 + c(t, t + 1, T i ). The conditional probability density of returns can be derived by summing over the number of jumps f t R i,t+1 = j=1 f t (R i,t+1 n t+1 = j) Pr t (n t+1 = j), (4.2) where Pr t (n t+1 = j) = (h y,t+1 ) j exp ( h y,t+1 )/j! is the probability of having j jumps which is distributed as a Poisson counting process. We can therefore write the log likelihood function as the summation of the log likelihoods for all eight futures contracts 8 T 1 L Fut = ln (f t (R i,t+1 )) i=1 t=1 In estimation we assume that the conditional variance and the jump intensity are equal across maturities. This is clearly a simplifying assumption that will worsen the fit, but it is useful for the purpose of comparison with option-implied estimates. When implementing maximum likelihood estimation, the summation in (4.2) must be truncated. We truncate the summation at 5 jumps per day. We have experimented with increasing the truncation limit beyond 5 and found that our results are robust. (4.3) 12

13 Equations (2.9) and (2.1) indicate that we need to separately identify the two unobserved shocks z t+1 and y t+1 and filter the conditional variance h z,t+1 and the conditional jump intensity h y,t+1 which enter the likelihood function. The structure of the model allows us to do this using an analytical filter, which is discussed in CJO (212). Using this filter, calculating z t+1 and y t+1 is straightforward and very fast. It takes less than a second to filter 38,24 futures contracts and about two seconds to filter 283,653 option contracts using Matlab on a standard PC. 4.2 Estimation Results Table 2 presents the maximum likelihood parameter estimates for the GARCH benchmark model and the four proposed jump models. The results are obtained using all eight futures contracts jointly in estimation for the time period For each jump model, we separate the parameters into two columns. The parameters with subscript y are reported in the column labeled Jump. The parameters with subscript z are reported in the column labeled Normal. Under each parameter estimate, we report its standard error calculated using the Hessian matrix. Under Properties, we report the implied long-run risk premiums for the normal and jump components, the percent of total variance captured by the normal and the jump component, the average annual volatility, the expected number of jumps per year implied by the parameter estimates, and the log likelihood. Some of these properties are discussed in more detail in Section 4.3 below. The log-likelihood values of the DVCJ, DVDJ, DVSDJ models are much higher than that of the GARCH model. This is not the case for the CVDJ model, which does not nest the GARCH model. To examine whether the jump models significantly improve over the GARCH model, we test the null hypothesis of no jumps. To implement this test, we use the standardized likelihood ratio test proposed by Hansen (1992, 1994). A likelihood ratio test of the null hypothesis of no jumps does not have the usual limiting chi-squared distribution because the jump parameters are unidentified under the null. Hansen s test is able to provide an upper bound to the asymptotic distribution of standardized likelihood ratio statistics, even when conventional regularity conditions (for example due to unidentified parameters) are violated. We calculate Hansen s test for each of the DVCJ, DVDJ, DVSDJ models compared with GARCH and report the standardized likelihood ratio and the corresponding simulated critical values in Table 3. Using Hansen s standardized LR test, we find that all three dynamic volatility jump models significantly improve over the GARCH model, suggesting that the null hypothesis of no jumps is rejected. Among the three jump models, the test result for the DVDJ model provides the strongest evidence against 13

14 the null. These statistical tests strongly suggest that incorporating jumps in addition to dynamic volatility helps to improve model performance. The CVDJ model, which has constant conditional variance of the normal component, performs the worst among the four jump models, with the lowest log-likelihood. This suggests that this extended Merton (1976) jump model, which has constant volatility and a GARCH dynamic for the jump intensity is not able to adequately fit the data, which confirms the importance of allowing for time-varying volatility in models for commodity futures. In the CVDJ model, it is not surprising that the jump component of CVDJ accounts for 63% of the total return variance, because the time variation in the return innovation is restricted to the jump intensity. The jump component is relatively more important for this specification because jumps are the only source of heteroskedasticity. In summary, the results for the CVDJ model suggest that jumps are a useful modeling tool, but that modeling time-varying volatility is at least as important. This finding is very similar to results obtained in equity and equity index markets. The DVCJ model improves model fit significantly by adding a simple constant jump component. For this specification, the average expected number of jumps is given by E h y,t+1 = w y. The estimate of.13 implies that jumps arrive at an annual frequency of.13x252, or approximately 26. This number is much higher than what is usually found in equity index markets, see the summary table in Broadie, Chernov, and Johannes (27). Most existing estimates in equity index markets find between one and three jumps per year. This suggests that jumps arrive more frequently in the crude oil market. Part of the explanation is that the models identify smaller jumps in crude oil futures as opposed to equity index returns. The estimate of the average jump size θ in the DVCJ model in Table 2 is -.69, which is approximately one-fourth of the average jump size in index returns documented in CJO (212) for a model with constant jump size. The results for the DVDJ model indicate that allowing for state-dependent jump intensities can further improve model performance. The estimate of k is statistically significant, confirming that the arrival rate of jumps depends on the level of volatility. The mean jump size in the DVDJ model is smaller (in absolute value) than in the DVCJ model and jumps arrive more frequently, with on average 27 jumps per year. The MLE estimates of the DVSDJ model indicate that there are on average 183 jumps per year. The likelihood for the most general specification, DVSDJ model, further improves on DVDJ. We run a likelihood ratio test to compare DVSDJ with CVDJ and the result is in favor of the unrestricted DVSDJ model. 14

15 The results for the DVDJ and DVSDJ models indicate that allowing for time-varying jump intensities can greatly improve model performance. The average jump arrival frequency for the DVDJ and DVSDJ models is significantly lower than for the DVCJ model, but other important model features are similar to the DVCJ model. 4.3 Model Implications We now further discuss the model properties listed at the bottom of Table 2. We report the decomposition of the total unconditional return variance into the normal and jump components. The total unconditional return variance, σ 2, is given by σ 2 σ 2 z + (θ 2 + δδ 2 2 )σ y (4.4) 2 2 where σ z and σ y are computed as the time series averages of h z,t+1 and h y,t+1. We report the normal contribution and jump contribution to the total return variance in percentages. Overall, the contribution of jumps to the total return variance is very high for the CVDJ model, since jumps are the only source of heteroskedasticity. For the other three models, the DVDJ and DVSDJ models have higher percentages (43.6% and 4.36%) for the jump contribution compared with the DVCJ, which has 15.48% of total variance contributed by the jump component. While the likelihood ratio tests indicate that allowing for time-varying jump intensities is supported statistically, these findings suggest that this model feature is also economically important. The average variance is very similar across models. The left panels in Figure 4 depict the time path of the conditional variance and clearly indicate that in fact the entire time path is very similar across models, with the exception of the CVDJ model. However, there is one important difference between the models with time-varying intensity (DVDJ and DVSDJ) and the DVCJ model. When the variance increases in 1991 and 28, the increase is more pronounced for the DVCJ model. The reason for this can be seen in the panels on the right in Figure 4. For the DVDJ and DVSDJ models, there is also a sharp increase in the jump intensity during those crisis periods. The paths of the jump intensity for the DVDJ and DVSDJ models are similar, but the intensity is more variable for the DVSDJ model. The variations in jump intensities affect the risk premiums, which are depicted in Figure 5. For the DVCJ model in the first row, the risk premium in the middle panel is constant, and all the time variation in the total risk premium in the right panel is due to variation in the normal risk premium. In the case of the DVDJ and DVSDJ models, a large amount of the increase in the total risk premium in 1991 and 28 15

16 is due to the increase in the jump risk premium. This effect is stronger in the more richly parameterized DVSDJ model. For the CVDJ model in the second row, all the time variation in risk premiums is due to the jump risk premium. This clearly does not lead to as much time variation in total risk premiums as for the other models. Overall, Figure 5 clearly indicates that jump risk premiums are economically important, and that they represent a significant component of the total risk premium. Under Properties in Table 2 we also report the averages of the risk premiums over the sample. On average the jump risk premium is large. For the CVDJ model, it is on average larger than the risk premium for the normal component, but this is of course due to the model s restrictive structure. It is also noteworthy that the average risk premiums are very similar for the DVCJ and DVSDJ models, which have a very different structure. The jump risk premium is relatively smaller in the case of the DVDJ model. Figure 6 applies the analytical filter to decompose futures returns in the jump and normal component to infer their relative importance. The middle column contains the filtered jump component and the right column contains the filtered normal component. The figures clearly indicate that most of the time, the normal component dominates return innovations. However, in crises, such as the first Gulf War in late 1991, the jump component explains more of the movement in returns than the normal component. The heteroskedasticity in the normal component is also apparent. The left columns depict the filtered number of jumps n t occurring each day. We find strong evidence for multiple jumps per day in all jump models, especially during the first Gulf War, when the DVDJ and DVSDJ models indicate 1 jumps per day. For the purpose of option valuation, the time path of the conditional variance is of paramount importance. However, different models often yield variance paths that are nearly similar, as evidenced by Figure 4. It is therefore of great interest to inspect differences in the conditional third and fourth moments. We now turn to this evidence. The first four conditional moments are given by: E t (R t+1 ) μ t+1 = r t+1 + λ z 1 2 h z,t+1 + λ y ξ h y,t+1 + c(t, t + 1, T) (4.5) Var t (R t+1 ) = h z,t+1 + (θ 2 + δδ 2 )h y,t+1 (4.6) θ(θ 2 + 3δδ 2 )h y,t+1 Skew t (R t+1 ) = 3 h z,t+1 + (θ 2 + δδ 2 )h y,t+1 2 (4.7) 16

17 Kurt t (R t+1 ) = 3 + (θ4 + 6θ 2 δδ 2 + 3δδ 4 )h y,t+1 h z,t+1 + (θ 2 + δδ 2 )h y,t+1 2 (4.8) where Skew t (R t+1 ) and Kurt t (R t+1 ) are the conditional skewness and the conditional kurtosis of futures returns respectively. From equation (4.7), it is clear that in presence of jumps, when h y,t+1 is positive, the sign of the conditional skewness depends on the sign of the mean jump size θ. Both skewness and kurtosis are critically affected by the parameters θ and δδ. Figure 7 plots the conditional one day ahead skewness in (4.7) and kurtosis in (4.8) for each of the four jump models. The estimated average jump size θ is negative for all four jump models, and therefore the conditional one day ahead skewness is negative. However, skewness is rather small, and interestingly it is smaller for the more flexible models. Conditional kurtosis is of the same order of magnitude for the DVCJ, DVDJ, and DVSDJ models, but contains more sharp peaks for the DVCJ model. Clearly, for a given model outliers in model skewness and kurtosis are related, which is driven by the parameterization. 5. Option Valuation Theory for Crude Oil Futures We first characterize the risk-neutral dynamics. Subsequently we derive the closed-form option valuation formula. 5.1 The Equivalent Martingale Measure and Risk-Neutral Dynamics The estimates obtained from futures prices in Section 4 are physical parameters. To value crude oil options, we need return dynamics under the equivalent martingale or risk-neutral measure. In a framework with compound Poisson processes, the futures price can jump to an infinite set of values in a single period, and the equivalent martingale measure is therefore not unique. We proceed by specifying the conditional Radon-Nikodym derivative dq t+1 dp t+1 = dq t dp t exp (Λ z z t+1 + Λ y y t+1 ) E t [exp(λ z z t+1 + Λ y y t+1 )] (5.1) This Radon-Nikodym derivative specifies a risk-neutral probability measure if and only if Λ z and Λ y are determined by 17

18 Λ z + λ z = (5.2) λ y e θ+δ2 2 1 e Λ yθ+ Λ 2 y δ e ( y+.5)δ 2 +θ = (5.3) The futures return process under the risk-neutral dynamic takes the form log F t+1,t F t,t = r t h z,t+1 ξ y (1) h y,t+1 with the following variance and jump intensity dynamics + z t+1 + y t+1 + c(t, t + 1, T) (5.4) h z,t+1 = ω z + b z h z,t + a z h z,t (z t c z h z,t ) 2 + d z y t (5.5) h y,t+1 = ω y + b y h y,t + a y h z,t (z t c y h z,t ) 2 + d y y t (5.6) where h y,t+1 = h y,t+1 Π, Π = e Λ yθ+ Λ y 2 δ 2 2, θ = θ + Λ y δδ 2, ξ(1) = e θ + δ2 2 1, Λ z = λ z, ω y = ω y Π, a y = Πa y, c z = c z Λ z, c y = c y Λ z, d y = Πd y, and y t+1 ~J(h y,t+1, θ, δδ 2 ). The risk neutral dynamic for the GARCH benchmark model in Section 2.1 is a special case of (5.4)-(5.6) with h y,t+1 = y t+1 = 5.2 Closed-form Option Valuation Under the risk-neutral measure, the generating function for the asset process in (5.4)-(5.6) takes the following form f(φ; t, T) E t Q F T φ = F φ t,t exp(a1(φ; t, T) + B1(φ; t, T)h z,t+1 + C1(φ; t, T)h y,t+1 ) Here we present the analytical solutions to the affine coefficients A1(φ; t, T), B1(φ; t, T), and C1(φ; t, T). Details on the derivation are provided in Appendix B. A1(φ; t, T) = φr t+1 + A1(φ; t + 1, T)+B1(φ; t + 1, T)ω z + C1(φ; t + 1, T)ω y 1 2 log(1 2B1(φ; t + 1, T)a z 2C1(φ; t + 1, T)a y ) (5.7) (5.8) 18

19 B1(φ; t, T) = φμ 1z + B1(φ; t + 1, T)(b z + a z c z 2 ) + C1(φ; t + 1, T)a y c y 2 + φ 2B1(φ; t + 1, T)a zc z 2C1(φ; t + 1, T)a y c y 2 2(1 2B1(φ; t + 1, T)a z 2C1(φ; t + 1, T)a y ) (5.9) C1(φ; t, T) = b y C1(φ; t + 1, T) + φμ 1y + ξ y (Ξ) (5.1) where μ 1z = 1 2, μ 1y = ξ y (1), ξ y (Ξ) = e θ Ξ Ξ y 2 δ 2 1, with Ξ = φ + B1(φ; t + 1, T)d z + C1(φ; t + 1, T)d y (5.11) By imposing the restrictions h y,t+1 GARCH benchmark model in Section 2.1. = ω y = a y =, (5.7)-(5.11) reduces to the generating function for With the risk neutral generating function (5.7), we can value European options using the Fourier inversion method as in Heston (1993), Heston and Nandi (2), and Duffie, Pan and Singleton (2). The price of a European call option on a futures contract is given by CO(t, T co, T, K) = E Q t e t Tco r(s)ds (F(T co, T) K) + = F(t, T) π Re f (iφ + 1) K iφ iφf dφ (1) e Tco t r(s)ds K( π Re f (iφ) K iφ d φ) iφ (5.12) where CO t, T co, T, K is the time t price of a European call option expiring at time T co with strike K on a futures contract expiring at time T, and K is the strike price. 6. Estimation using Option Data We first discuss the loss function used to estimate the model parameters from option data. Subsequently we discuss the empirical results. 6.1 MLE using Option Data It is possible to use the parameter estimates in Table 2 obtained through MLE estimation on futures data to compute option prices using the option valuation formulae. However, this procedure exclusively uses 19

20 historical information and ignores the expectations about the future evolution of the futures prices that are embedded in option prices. We therefore study the models option valuation performance by maximizing the log-likelihood function for the option contracts to match model option values as closely as possible to observed market prices. We use a loss function based on implied volatilities, inverting option prices into implied volatilities. This results in a cross-section of market data that is of similar magnitude along the moneyness, maturity, and time-series dimensions. Option prices differ significantly along these dimensions. Define the model error u k = σ k,t σ k,t (O k,t (h t (θ ))) (6.1) where σ k,t is the Black (1976) implied volatility of the k th observed option price at time t, and σ k,t (O k,t (h t (θ ))) is the implied volatility converted from each computed option price, O k,t (h t (θ )), using the Black (1976) formula. Assuming normality of the implied volatility errors, the log-likelihood function based on options is L Opt = N 2 ln(2πσ u 2 ) 1 2 u k 2 N σ2 u k=1 where N = 283,653 is the total number of option contracts. We estimate parameters based on options data only by maximizing L Opt. We report the optimized likelihoods as well as the corresponding volatility root mean squared error (IVRMSE) (6.2) IVRMSE = 1 T 1 N [σ k,t σ k,t (O k,t (h t (θ )))] t,k 2 (6.3) where T = 4,753 is the number of days used in our analysis. Ideally one would fit the model directly to implied Black volatilities. However, since the optimization routine requires computing implied volatility from model prices at every function evaluation, this approach is extremely slow. We follow TS (29) and fit the model to option prices scaled by their Black (1976) vega, that is, the sensitivities of the option prices to variations in log-normal volatilities. This approach is motivated by the approximation σ k,t O k,t /Ѵ k,t, where Ѵ k,t is the Black (1976) vega associated with the k th observed option price at time t. This approximation has been shown to work well in existing work. Thanks to the use of the analytical filter, the quasi-closed form option valuation formula, and the use of the vega-scaled prices, the optimization problem is feasible with our large data sets. 2

21 6.2 Empirical Results Table 4 reports the MLE estimates on crude oil options for the benchmark GARCH model and the four jump models. We report the log likelihood and the IVRMSE for each model at the bottom of the table, as well as a number of important model features implied by the parameter estimates. Note that unlike the parameters in Table 2, these are risk-neutral parameters. Similar to the estimation results based on crude oil futures data in Table 2, the log-likelihood values for the DVCJ, DVDJ, and DVSDJ models are much higher than that of the GARCH model. The CVDJ model performs the worst among the four jump models. We find that jump models, except for CVDJ model, provide substantial improvements in option pricing performance. Jump models with time-varying intensities, DVDJ and DVSDJ, obtain IVRMSEs of 6.83 and 6.75 respectively and outperform the GARCH model by approximately 6% and 7%. This result further confirms that incorporating timevarying jumps in addition to the dynamic volatilities helps to improve model fit. The most striking difference between these estimates and the estimates based on futures returns in Table 2 is that the average jump size θ implied by option prices is much larger, and that jumps occur less frequently. These findings suggest that jumps implied by crude oil options are larger and rarer events comparing with those estimated using crude oil futures returns. 7. Estimation using Futures and Options Data The parameters obtained by minimizing option errors in Section 6 are of substantial interest, and the resulting differences in option fit are a useful indicator of the models performance. It must also be noted that the exercise in Section 6 does not ignore the underlying futures returns, because volatilities and jump frequencies are filtered from the futures returns. However, the futures returns do not directly enter the loss function. So while this exercise imposes considerable discipline upon the models, it has nevertheless two important drawbacks. First, if a model is richly parameterized, only fitting the option data may result in overfitting. Second, the price of risk parameters, which are some of the most economically important model parameters, cannot be reliably identified using option data only. We therefore follow Bates (1996), who suggests that the most stringent test of an option pricing model lies in its ability to jointly fit the option data and the underlying returns. In our case, this means 21

22 that we have to construct a loss function that contains a crude oil option component as well as a futures return component. We first explain how we combine the option data with the underlying futures data to conduct a joint likelihood. We then discuss the parameter estimates, and subsequently we compare the most important model properties with the properties implied by the physical parameters from Table The Joint Log Likelihood Function In previous sections we used MLE to estimate the jump models using futures returns and option implied volatilities respectively. The log-likelihood functions for futures and options are defined in equations (4.3) and (6.2). To jointly estimate the jump models using the option data and the underlying futures data, we maximize the weighted average of the log-likelihoods of futures and options. The number of option contracts in the data set is much larger than the number of futures contracts. To ensure that joint parameter estimates are not dominated by options, we assign equal weight to each log-likelihood. The resulting weighted joint log-likelihood is L Joint = M + N L Fut 2 M M + N L Opt + 2 N where M=38,24 is the total number of futures contracts and N=283,653 is the total number of option contracts. The joint likelihood function can then be obtained using the results in Sections 4.1 and 6.1. (7.1) 7.2 Empirical Estimates and Model Implications Table 5 contains the parameter estimates obtained by maximizing the joint log-likelihood function in (7.1) for the GARCH model and four jump models. At the bottom of the table, we also report the IVRMSE and model properties implied by the parameters such as the long-run risk premium, the percent of total annual variance explained by the normal and the jump component, the average annual volatility, and the expected number of jumps per year. Similar to the results based on options and the results based on futures, the jump models, except for the CVDJ model, outperform the benchmark GARCH model according to both log-likelihood values and IVRMSE values. The most richly parameterized jump model, DVSDJ, performs best. The most important conclusion from the parameter estimates is that the structure of the estimated jump processes is very different from the jumps estimated from futures returns in Table 2. Results are more similar to those based on options in Table 4, in the sense that jumps are larger and more infrequent. 22